This project is one of a number that are in competition for funding from the NERC Great Western Four+ Doctoral Training Partnership (GW4+ DTP). The GW4+ DTP consists of the Great Western Four alliance of the University of Bath, University of Bristol, Cardiff University and the University of Exeter plus five Research Organisation partners: British Antarctic Survey, British Geological Survey, Centre for Ecology and Hydrology, the Natural History Museum and Plymouth Marine Laboratory. The partnership aims to provide a broad training in earth and environmental sciences, designed to train tomorrow’s leaders in earth and environmental science http://nercgw4plus.ac.uk/
Project Background
Marine and freshwater plankton are a physiologically and morphologically diverse group of organisms that inhabit aquatic environments around the world. Many plankton are microscopic and are invisible to the naked eye. However, in oceans and bodies of freshwater their population sizes can grow to levels that allow them to be easily viewed from space. In the ocean, microscopic photosynthetic plankton perform a similar role to terrestrial plants, trapping energy from the sun and using it to form organic material. In turn, these organisms are grazed by different types of microscopic zooplankton that feed heterotrophically, and ultimately support the growth of organisms higher up the food chain including fish.
To build an understanding of how marine planktonic communities are structured, how they function and how they change in time, it is important to be able to monitor the abundance of different planktonic organisms. As the manual identification and quantification of plankton is both time consuming and costly, a significant amount of effort has gone into developing automatic imaging and classification systems using cutting edge machine learning techniques [1, 2]. However, to date these techniques have relied on the existence of large libraries of labelled images, which themselves are difficult to create and rely on the work of expert taxonomists [3]. If a high accuracy, low-shot classification system for the automatic recognition of plankton images could be developed, it would have significant advantages over existing technologies.
With the further development of deep learning, computer vision has made great progress, mainly due to the powerful feature extraction capabilities of Convolutional Neural Networks (CNN) and the creation of large-scale data sets used for training, such as ImageNet [4]. However, a CNN’s capability to recognize objects is greatly reduced if only a small training sample size is available. For humans, only a tiny number of samples are required for the successful identification of objects [5]. To make machines also able to recognise objects with a small training sample size, the field of low-shot learning has been gradually and continuously developing [6,9,10].
Project Aims and Methods
At present, most algorithms for classifying marine plankton images depend on the existence of numerous training samples, and mainly CNN and transfer learning methods are adopted to train image classifiers [1,2]. The aim of this project is to investigate the effectiveness of low-shot machine learning techniques for the automatic classification of plankton image data when only limited training data is available. Specifically, we will focus on two different types of learning paradigms: (1) Few-shot and (2) Zero-shot:
1. Few-shot learning is designed to adapt quickly to new categories from few examples. The methods work by employing techniques for model initialization, metric learning, data enhancement, and transfer learning. In this project, we will focus on meta-learning-based models [6], which prevent the model from overfitting by extracting transferable knowledge from a set of tasks, resulting in a more generalized model.
2. Zero-shot learning is a particular problem setup in machine learning, where at test time, a learner observes samples from classes that were not observed during training and needs to predict the class they belong to [10]. We will consider the attribute annotations of plankton and attention based deep learning models [10] for mapping visual information from plankton images to the attribute space where the prediction of images from unseen classes can be done during the inference time.
The project will require the student to employ exciting and innovative techniques from the cutting edge of computer vision research to the important problem of plankton classification and enumeration. They will evaluate the accuracy of different classification algorithms; contrast them with current state of the art approaches; and investigate approaches for maximising performance with a view to operational deployment. The student will work with automatically acquired plankton image data that has been collected over multiple years by expert taxonomists at Plymouth Marine Laboratory (PML). To ensure the student has a solid understanding of how the image data is collected and the various scientific questions it is used to address, the student will be encouraged to take part in sampling work at sea aboard the Plymouth Quest, and to assist with laboratory-based classification in PML’s laboratories.
Candidate requirements
The project will suit a student with a degree in a numerate discipline (e.g., computer science, physics, mathematics) who has a strong background in computing, and in particular Python programming. A good working experience with deep learning frameworks, such as PyTorch or TensorFlow, is desirable.
Project partners
The student will have the exciting opportunity to work collaboratively with the University of Exeter, Plymouth Marine Laboratory, and the University of Glasgow. The student will be based in the University of Exeter and will have ample opportunities to visit other supervisors at their corresponding institutes.
Useful links
For information relating to the research project please contact the lead Supervisor Dr Anjan Dutta (webpage: http://emps.exeter.ac.uk/computer-science/staff/ad735) via email: [Email Address Removed]
How to apply
In order to formally apply for the PhD Project you will need to go to the following web page.
https://www.exeter.ac.uk/study/funding/award/?id=4239
The closing date for applications is 1600 hours GMT on Friday 10th January 2022.
Interviews will be held between 28th February and 4th March 2022.
If you have any general enquiries about the application process please email [Email Address Removed] or phone: 0300 555 60 60 (UK callers) or +44 (0) 1392 723044 (EU/International callers). Project-specific queries should be directed to the main supervisor
Continue with Facebook
